# Claude Opus 5.5: Anthropic's model for coding agents and knowledge work

> What Claude Opus 5.5 is, which benchmarks Anthropic published and how to read them, what it costs ($4/$20) and when to pick it. Data as of November 15, 2026.

*Tags: ai, models, anthropic, coding, agents*

---


> [!info] Quick Definition
> **Claude Opus 5.5** is a closed language model from Anthropic, released on September 22, 2026 as the first in the Claude 5.5 family. It is built for long-running coding agents and knowledge work, and Anthropic's documentation recommends it as the starting point for most tasks.


## What is Claude Opus 5.5 and when did it come out?

It is the upper-middle model of the family: above Sonnet 5.5 and [[claude-haiku-5-5|Haiku 5.5]] on price, and below [[claude-fable-5-2|Fable 5.2]]. Anthropic's model guide says that if you are unsure which to use, you should start with Opus 5.5 for most workloads. It thinks before answering using adaptive thinking that cannot be switched off (see [[reasoning-models|reasoning models]]).

| Fact | Value |
|---|---|
| API ID | `claude-opus-5-5` |
| Release | September 22, 2026 |
| Context and output | 1M tokens of context; up to 128K of output |
| Price | $4 input / $20 output, per million tokens |
| Input and output | Text and images in, text out |
| Knowledge cutoff | June 2026 |
| Default effort | `medium` |
| Open or closed | Closed: used through the API and cloud platforms |

## What is it for?

The documentation describes it for autonomous coding agents that work for hours, large refactors, complex systems engineering, vision-heavy workflows and computer use. It also fits knowledge work: reading a lot of material, reasoning and delivering a result.

For people who design and code with AI, it fits agents that edit a repository; to see how one works, check [[claude-code|Claude Code]]. It can also review screenshots, because it accepts images. If your use is conversation and writing, see [[claude-for-designers|Claude for designers]].

## Which benchmarks did it publish and how do you read them?

Anthropic reports these scores in the Opus 5.5 announcement, using its own test harness and `max` effort, except Terminal-Bench 4.0 (`xhigh`). They are not an independent measurement.

| Benchmark | Score | Measured by | What it measures, per the source |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% at `xhigh` effort (standard error of ±2.6 points) | Anthropic | How well it completes complex, multi-step professional tasks in a command-line interface |
| CursorBench 4.0 | 57.8% at `max` effort (52.5% at default effort, `medium`) | Anthropic | Coding agents on ambiguous, multi-file tasks taken from real Cursor sessions |
| FrontierCode v1.1 (Main) | 54.4% | Anthropic | Whether an agent's code changes would be merged |
| GDPval-AA v2.1 | 1846 Elo | Anthropic | Real-world professional work across 44 occupations |
| AutomationBench | 40.0% | Zapier (early-access evaluation, reported by Anthropic) | Whether an agent can carry out real business workflows across many connected apps (built by Zapier) |
| OSWorld 2.1 (partial) | 81.8% | Anthropic | How well an agent operates a real computer on long, multi-step tasks (description from the Haiku 5.5 page) |
| Humanity's Last Exam, with tools | 67.7% | Anthropic | Expert-level academic knowledge and reasoning (description from the Haiku 5.5 page) |

How to read them:

- **Look at the margin of error.** Anthropic gives ±2.6 points on Terminal-Bench 4.0: a gap of one or two points between models can be noise.
- **Look at the effort.** On CursorBench 4.0 there is about a 5-point gap between the default effort (`medium`) and `max`. A score without its configuration cannot be reproduced.
- **Compare within the same table.** In that announcement, Opus 5 scored 52.3% on Terminal-Bench 4.0 and Opus 5.5 scored 66.4%: same test, same lab.
- **Do not build rankings against other labs.** Each one runs its tests its own way.
- **Your case rules.** Measure with your own tasks; the guide is in [[ai-benchmarks|how to read AI benchmarks]].

## How much does it cost and how much context does it have?

As of November 15, 2026, Anthropic's documentation lists these prices per million tokens:

- Input: $4. Output: $20. Opus 5 cost $5 / $25.
- Cache read: $0.20. 5-minute cache write: $5.
- Batch processing (Batch API): 50% discount on input and output.
- Fast mode (up to 2.5 times faster generation): $8 input / $40 output.

The 1M-token context is about 555,000 words, according to the documentation. To understand the units, read [[tokens|tokens]], [[context-window|context window]] and [[cost-per-token|cost per token]].

Illustrative example of the math: a task with 200,000 input tokens and 20,000 output tokens costs $0.80 + $0.40 = $1.20, and half of that in batch. Real spend varies with what the model produces. With [[prompt-caching|prompt caching]] you lower the cost of what you repeat.

## When should you pick it and when not?

Pick it when:

- You want one default model for code and knowledge work and do not yet know whether you need more.
- An agent has to work for hours on a repository or a task with many files.
- The task combines text and images, such as reviewing screens or charts.

Do not pick it when:

- The task is high-volume and low-complexity. For classifying, extracting or routing, [[claude-haiku-5-5|Haiku 5.5]] costs a fraction.
- You already tried Opus 5.5 at high effort and it falls short: Anthropic then suggests [[claude-fable-5-2|Fable 5.2]].
- You need to turn reasoning off: on Opus 5.5 you cannot.

To see options from other labs, check [[current-ai-models|current AI models]].

## Example

Illustrative example: a product team wants an agent to migrate the buttons of a design system to new color tokens across 60 files. They give the agent the repo's instructions file and ask for a change in three files as a trial. They review the diff by hand and run the visual tests. If the result is correct, they launch the rest on a separate branch and review before accepting.

What to try tomorrow: take a small, repeatable coding task, run it with Opus 5.5 at `medium` effort and then at `high`, and compare the result, the time and the cost. That way you decide with your own data which effort suits you.

## Common mistakes

- Using Opus 5.5 to classify thousands of short texts, where Haiku 5.5 is much cheaper.
- Taking a benchmark figure at face value without checking the margin of error or the effort.
- Comparing Anthropic's scores with another lab's as if they were the same exam.
- Accepting an agent's change without reading the diff or running tests.
- Budgeting without checking prices again: they change.

## Sources

- [Anthropic: introducing Claude Opus 5.5](https://www.anthropic.com/news/claude-opus-5-5)
- [Documentation: Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/overview)
- [Documentation: Claude models overview](https://platform.claude.com/docs/en/about-claude/models/overview)
- [Documentation: pricing](https://platform.claude.com/docs/en/about-claude/pricing)
- [Anthropic: introducing Claude Haiku 5.5 (descriptions of OSWorld 2.1 and Humanity's Last Exam)](https://www.anthropic.com/news/claude-haiku-5-5)


---

Source: https://www.fernandoux.com/en/wiki/ai/claude-opus-5-5/
